Analysis of Liver Cancer Cell Lines Identifies Agents With Likely Efficacy Against Hepatocellular Carcinoma and Markers of Response

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Analysis of Liver Cancer Cell Lines Identifies Agents With Likely Efficacy Against Hepatocellular Carcinoma and Markers of Response Stefano Caruso, Anna-Line Calatayud, Jill Pilet, Tiziana La Bella, Samia Rekik, Sandrine Imbeaud, Eric Letouzé, Léa Meunier, Quentin Bayard, Nataliya Rohr-Udilova, et al.

Human tumor-derived cell lines have been widely used as models for studying cancer biology. They are very useful to understand mechanisms that drive resistance and sensitivity to anti-cancer compounds, in particular when access to tissue samples is limited as in HCC for which non-invasive imaging has replaced biopsy for diagnosis. 7 Over the past decade several large-scale pharmacogenomics studies in cancer cell lines including NCI-60, CCLE (Cancer Cell Line Encyclopedia) and GDSC (Genomics of Drug Sensitivity in Cancer) have proven their value for biomarker discovery as well as to uncover mechanism of drug action and determine molecular contexts associated with specific tumor vulnerabilities. 8

Cell lines
The 34 Liver cancer cell lines (LCCL) were collected from public repositories or collaborations and grown in monolayers in quite similar conditions, at 37°C in a humidified 5% CO 2 incubator (Supplementary Table 1). Cell line identity was verified by whole exome sequencing (WES), BEL7402 and SMMC7721 cell lines were excluded because contaminated by HeLa cells as well as SK-HEP-1 that has an endothelial origin. 11

Identification of putative somatic variants and copy number analysis
Identification of gene mutations and copy number alterations (CNA) was performed by WES (Supplementary Material). TERT promoter and exon 1 of ARID1A were screened by Sanger sequencing because of low coverage in WES, as previously described. 12

Selection of cancer driver genes in primary HCC tumors and comparison with LCCL
We selected 72 HCC cancer driver genes from 5 HCC publically available datasets to compare their alteration frequency between HCC and LCCL (Supplementary Table 2A

RNA-sequencing
Total mRNA extraction was performed for the 34 LCCL using RNeasy mini kit (Qiagen) and quality was checked. 5 µg of total RNA was used for sequencing (Supplementary Material).
We kept only fusions validated by BLAST and with at least 10 split-reads or pairs of reads spanning the fusion event, and we removed fusions identified at least twice in a cohort of 36 normal liver samples. HBV insertions were screened as described in the Supplementary Material.

Transcriptome analysis
Consensus clustering was performed with Bioconductor ConsensusClusterPlus package.
Principal component analysis using the first 3 components was also generated.

Single and combination drug screening
We analyzed 31 therapeutic compounds on the whole panel of 34 LCCL (Supplementary   Table 4). Drug screening was performed using the HP D300 digital dispenser (Tecan) to create dose-response curves as described in the Supplementary Material.

Identification of biomarkers related to drug sensitivity
To identify molecular features associated with drug response we performed elastic net regression as described in the Supplementary Material.

Statistical analysis
Statistical analysis and data visualization were performed using both R software version 3. using Pearson r correlation when both variables were normally distributed with the assumptions of linearity and homoscedasticity or Spearman's rank-order correlation. All tests were two-tailed and P-value < 0.05 was considered as significant.

Tumors
The LICA-FR and TCGA cohorts of HCC patients including respectively 156 and 319 tumors cases were previously described. 15,16 M A N U S C R I P T

Liver cancer cell lines retain the genomic alterations identified in HCC tumors
We qualified and analyzed a total of 34 liver cancer cell lines (LCCL) including 32 derived from HCC and 2 from hepatoblastoma (HepG2 and Huh6) (Supplementary Table 1) that were compared to 821 HCC primary tumors including our HCC cohort (n=235), 17 and 2 independent public datasets from Korea (n=231), 18 and mixed Asian/European origin (n=355). 16  14 classified in the "proliferation class". 19 In contrast, CTNNB1 mutations belonging the "nonproliferation class" were less frequent in LCCL (15%) compared with primary tumors (29%).
CNA analysis identified recurrent homozygous deletions of CDKN2A/MTAP and AXIN1, and focal amplification containing CCND1 and FGF19 in LCCL as in HCC (Supplementary Figure 1B). As previously described, in primary HCC major associations between risk factors and gene mutation were found between alcohol intake and CTNNB1 and TERT promoter mutations and HBV infection and TP53 mutations. In LCCL, we confirmed the significant association between HBV infection and TP53 mutations (Supplementary Figure 3). In addition, 300 fusion transcripts were identified by RNA sequencing across the 34 LCCL.
Among them, 51 involved cancer driver genes related to HCC (n=11) or to other cancer types (n=40) (Supplementary Table 6). We also detected 33 chimeric HBV-human fusion transcripts in 13 LCCL with recurrent insertions in TERT promoter (3/34), as previously reported in HCC primary tumors (Supplementary Table 7). 16,20 Overall, our panel of LCCL Hoshida (S1-S2) corresponding to the "proliferation class" ( Table 10). Accordingly, miR-122-5p was reported as liverspecific and represents the most abundant miRNA in mature hepatocytes. 23 In the same line, miR-194-5p was described as a liver epithelial cell marker and its downregulation increased EMT and HCC metastasis in preclinical models. 24 The 126 proteins analyzed by RPPA included 82 total proteins and 44 phospho-proteins involved in various signaling pathways (Supplementary Table 3   HER3, FGFR4) that were dramatically downregulated in CL3 subgroup and expression of the hepatic progenitor marker cytokeratin 19 that was higher in CL1 and CL2 subgroups ( Figure   3B). Protein expression identified a more pronounced activation of the TGFß (TGFß-I-III and phospho-SMAD2/3) and the non-canonical ß-catenin pathways (phosphoSer675) in CL2 and CL3 subgroups ( Figure 3B).
We also identified two major networks of protein co-regulation in the whole panel of LCCL with proteins involved in DNA repair, cell cycle, apoptosis and in the PI3K/AKT/mTOR pathway ( Figure 3C). Finally, investigating relationship between protein expression and the mutational status of the HCC driver genes mutated in LCCL yielded 268 significant associations (Supplementary Table 13). As expected, cyclin D1 was overexpressed in LCCL harboring co-amplification of CCND1 and FGF19. AXIN1 protein was overexpressed in the In addition, consistent with their inactivation, mutations in the tumor suppressor genes (TSG) TSC2 and AXIN1 were associated with the downregulation of the corresponding proteins, we also found a known association between inactivating mutations of KEAP1 and overexpression of NQO1 protein ( Figure 4D).

Drug screening and molecular features associated with drug sensitivity
In our panel of 34 LCCL we screened 31 drugs including compounds approved or in clinical development in HCC or other cancers and targeting key pathways of liver tumorigenesis ( Figure 4A and Supplementary Table 4 and 14). We also analyzed four drug combinations with sorafenib including the AKT inhibitor MK-2206, the HDAC inhibitor resminostat and the two MEK1/2 inhibitors trametinib and refametinib.
We showed that the most potent drugs were those that target general processes, such as proteasome, mitosis or protein folding ( Figure 4A). Of note inhibitors targeting both PI3K/mTOR or mTOR alone were among the 7 most effective drugs with a median AUC close to doxorubicin, a chemotherapeutic agent used to treat HCC by transarterial chemoembolization ( Figure 4A). showed common sensitivity profiles for drugs with similar mechanism of action and identified two main subgroups of LCCL associated with transcriptomic subgroups (P<0.009) ( Figure   4B). Accordingly, the global drug response rate was higher in the most differentiated CL1 subgroup, compared with CL2 and CL3 subgroups that were less differentiated and resistant to most of the analyzed compounds ( Figure 4C)  Table 14). Refametinib (anti-MEK) and tanespimycin (anti-HSP90) were more efficient in CL1 compared with CL3 subgroup, while trametinib (anti-MEK1/2) showed higher efficiency both in CL1 and CL2 subgroups. Combination of sorafenib either with trametinib or refametinib was more potent both in CL1 and CL2 subgroups. We also identified 312 drug-protein predictive pairs ( Figure 4E and Supplementary Table 15). Among them, we validated the association between the high expression level of NQO1 and the high sensitivity to the two HSP90 inhibitors, alvespimycin and tanespimycin. 9,10,26 Expression of cytokeratin 19 was strongly associated with higher sensitivity to dasatinib (src-inhibitor), in agreement with the higher dasatinib vulnerability in a "progenitor-like" subtype of LCCL. 27 In our study, cytokeratin 19 expression was also associated with higher response to trametinib (anti-MEK1/2) and navitoclax (anti Bcl-2, Bcl-XL, and Bcl-w).
We also identified 143 significant associations between genetic alterations and drug sensitivity (Supplementary Tables 16 and 17) among which, TSC1/TSC2 inactivating mutations were linked with a higher sensitivity to the mTOR inhibitor rapamycin ( Figure 4F).
In addition, we confirmed in our large panel of LCCL, the hypersensitivity to the AURKA inhibitor Alisertib in the TP53-mutated cell lines ( Figure 4F). 28 Interestingly, the only METamplified cell line (MHCC97H, Figure 4D) was highly sensitivity to the two selective MET inhibitors (PHA-665752 and JNJ-38877605) as well as cabozantinib ( Figure 4D and Supplementary Table 14).
In order to explore among all the molecular features (genomic alterations, miRNA and mRNA expression), those that were the most associated with drug response, we used elastic net ( Table 18).

Sensitivity to the MEK inhibitor trametinib is related to RAS-MAPK genomic alterations and cell differentiation
We focused our analysis on the MEK inhibitor trametinib, that showed a bimodal sensitivity pattern, with a group of highly sensitive and a group of resistant LCCL ( Figure   5A). We identified oncogenic alterations known to activate the RAS-MAPK pathway in half of the LCCL ( Figure 5A). All these genomic alterations were mutually exclusive, they  Figure 5C). Accordingly, within the FGF19amplified cell lines, we identified a strong correlation between trametinib sensitivity and FGF19 mRNA level ( Figure 5D). A similar association was observed with the two FGFR4 inhibitors BLU-9931 and H3B-6527, which reinforce the link between trametinib sensitivity, FGF19 amplification and expression ( Figure 5D). Altogether, these results revealed that FGF19 amplification solely is not sufficient to predict sensitivity to both trametinib and FGFR4 inhibitors but the expression of a full pathway modulated by the context of differentiation is required.
Then, we searched for robust predictors of trametinib response among all the molecular features analyzed using elastic net regression. We identified 5 genes (HSD17B7, RORC, MRPS14, SERINC2, LAD1) with a high mRNA expression associated with higher trametinib sensitivity, named "trametinib 5 gene-score" to predict accurately the response ( Figure 5E and Supplementary Table 19). We validated these results in the GDSC external dataset Overall, G1/S2 subclasses of HCC may be the best candidate for a trametinib or anti-FGFR4 therapy while G3/S1 subclasses are unlikely to respond.

DISCUSSION
In the present study, we showed that our panel of LCCL recapitulates the diversity of the most aggressive "proliferation class" of HCC both at the genomic and transcriptomic levels and this is a unique tool to translate our understanding of liver cancer development into therapeutics (summarized in Figure 7).
By combining genomic, transcriptomic and protein profile analysis in our panel of 34 LCCL, we identified strong similarities with the established HCC molecular subclasses, suggesting that LCCL are representative models of primary tumors. We identified three robust transcriptomic subgroups of LCCL driven by the differentiation state and sharing features similar to those described in HCC tumors. The CL1 "hepatoblast-like" subgroup of cell lines expresses hepato-specific genes and fetal/progenitor markers, it corresponds to the "progenitor subclass" of HCC. 19 CL2 "mixed epithelial-mesenchymal" and CL3 "mesenchymal-like" subgroups were less differentiated with an activation of the TGFß and non-canonical ß-catenin pathways, they were more similar to the "Wnt-TGFß" 22  Our findings provided novel insights regarding the crucial interplay between the differentiation context, the genetic alterations and drug response in HCC (summarized in Figure 7). Strikingly, the global drug response rate among LCCL was related to the transcriptomic subgroup and the cell differentiation state with the most differentiated CL1 subgroup showing the highest drug sensitivity.
Cell differentiation also interferes with specific signaling pathways. inhibitors, which extends other previous findings. [30][31][32] Remarkably, this group of LCCL was also highly sensitive to inhibition of MEK1/2, an effector downstream FGF19/FGFR4, with trametinib, which is already approved in BRAF-mutated melanoma. 33 In LCCL, trametinib has demonstrated higher potency than the anti-FGFR4 BLU-9931 thereby representing a new attractive drug for targeting HCC addicted to the FGF19/FGFR4 pathway. Accumulating evidence indicate that the differentiation context plays a determinant role in treatment response, in particular it was shown that a therapy targeting a specific mutation will not necessarily have the same efficacy in tumors sharing the same mutation but arising in different tissue types. 34 In this line, a recent report demonstrated that gene expression and the tissue of origin predicted much better drug sensitivity in pan-cancer cell line analysis than genetic alterations. 35 Here, we showed that this concept could be also generalized to tumors from the same organ. Interestingly, trametinib is also efficient in LCCL harboring other alterations in the RAS-MAPK pathway such as RPS6KA3 inactivation or MET amplification.
However, MET amplifications, even if they are rare events, are more efficiently targeted by specific MET inhibitors including cabozantinib and these findings could be translated in the clinics as we confirmed recently MET oncogenic addiction in a patient with an advanced HCC amplified for MET, who achieved a complete tumor response after treatment by teponinib, a specific Met inhibitor. 36 Our study has also highlighted a synergistic effect of sorafenib in combination with MEK1/2 inhibitors with higher sensitivity in the CL1 and CL2 subgroups of LCCL, in lines with recent studies in HCC preclinical models and in HCC patients. 37,38 Our screening also identified potential new attractive drugs already approved in the clinics in specific molecular contexts. Our results showed responses in LCCL harboring inactivating mutations in TSC1 or TSC2 treated by an mTOR inhibitor, suggesting that the 7% of HCC demonstrating the same alterations could benefit from rapamycin or alternative inhibitors, in line with previous reports (Figure 7). 16,18,[39][40][41] Dasatinib also showed enhanced efficiency in LCCL expressing high levels of cytokeratin 19. 27 We recently reported a specific enrichment of immunohistochemical expression of CK19 in the "progenitor subclass" of HCC which may represent a good candidate for dasatinib therapy. 41 Additionally, we identified other potential drugs that may specifically target the "progenitor subclass" of HCC such as linsitinib, an inhibitor of IGF1R, or sorafenib in combination with the anti-AKT MK-2204. Indeed, linsitinib hypersensitivity in the CL1 subgroup of LCCL, that recapitulates the "progenitor subclass" of HCC, is consistent with the strong overexpression of IGF2. Accordingly, a recent work in transgenic mice demonstrated the pro-oncogenic role of IGF2 in the liver and showed that blocking IGF2 by an antibody efficiently impaired growth of liver tumor cells overexpressing IGF2, in vitro and in vivo. 42 In the present work, we also enlighten the specific vulnerability of TP53-mutated LCCL to the AURKA inhibitor alisertib corroborating a recent study in mice. 28  anti-angiogenic activity that could not be explored in vitro and represents a limitation in the use of cellular models.
In conclusion, our work showed that LCCL represent a valuable and powerful resource for drug-biomarker discovery that may be useful to guide future clinical trials.       Table 19). On the right below, correlation between trametinib response and the mean expression of the 5 mRNA (green) overexpressed in sensitive LCCL ("trametinib 5-gene score") (Spearman's test).
Variance stabilized values were used for mRNA level except for MET and ERBB4.

M A N U S C R I P T
A C C E P T E D ACCEPTED MANUSCRIPT Huh6 HepaRG JHH1